This research paper reports preliminary results of data-driven modeling of segmentalphoneme duration for Marathi. Classification and Regression Tree based data driven duration modeling for segmental duration prediction is presented. A number of features are considered and their usefulness and relative contribution for segmental duration prediction is assessed. Objective evaluation of the duration model, by root mean squared prediction error and correlation between actual and predicted durations, is performed.

1. Sangramsing Kayte Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 5, Issue 11, (Part - 4) November 2015, pp.115-119
www.ijera.com 115 | P a g e
Duration for Classification and Regression Treefor Marathi Text-
to-Speech Synthesis System
Sangramsing N. Kayte1
, Monica Mundada1
,Dr. Charansing N. Kayte
2
, Dr.
Bharti Gawali*
1,3
Department of Computer Science and Information Technology Dr. Babasaheb Ambedkar Marathwada
University, Aurangabad
2
Department of Digital and Cyber Forensic, Aurangabad, Maharashtra
Abstract
This research paper reports preliminary results of data-driven modeling of segmentalphoneme duration for
Marathi. Classification and Regression Tree based data driven duration modeling for segmental duration
prediction is presented. A number of features are considered and their usefulness and relative contribution for
segmental duration prediction is assessed. Objective evaluation of the duration model, by root mean squared
prediction error and correlation between actual and predicted durations, is performed.
I. Introduction
Accurate estimation of segmental durations is
crucial for natural sounding text-to-speech synthesis
[1-2]. Variation in segmental duration serves as a cue
to the identity of a speech sound and helps to segment
a continuous flow of sounds into words and phrases
thereby increasing the naturalness and intelligibility.
The primary goal in duration modeling is to model
the duration pattern of natural speech, considering
various features that affect the pattern. An important
restriction being that, due to the nature of the Text-to-
Speech synthesis problem, i.e., as only text is
provided for the synthesis, only those features that
can be automatically derived from text can be
considered.
The approaches to segmental duration modeling
can be divided into two categories: rule-based and
corpusbased. The most prevalent rule-based duration
model is a sequential rule based system proposed by
Klatt [2], which is implemented in the MI-Talk
classification [3]. In this system, starting fromsome
intrinsic rule, the durationof a segment is modified by
rules that are applied sequentially. Models of this type
have been developed for several languages [4, 5].
However, rule-based models often over-generalize
and cannot handle exceptions well without getting
exceedingly complicated. When large speech corpora
and the computational means for analyzing these
corpora became available, new data-driven
approaches based on Classification and Regression
Trees [6], linear statistical models and Artificial
Neural Networks [7] have been increasingly used for
duration modeling.
In this paper, duration modeling for Marathi is
performed using data-drivenapproachbased on CART.
Classification and Regression Trees are models based
on self-learning procedures that sort the instances in
the learning data by binary questions about the
attributes that the instances have. It starts at the root
node and continues to ask questions about the
attributes of the instance down the tree until a leaf
node is reached [1,2,3,8]. For each node, the decision
tree algorithm selects the best attribute, and also the
question to be asked about that attribute. The
selection is based on what attribute and question
about it divide the learning data so that it gives the
best predictive value for right classification. CART
modeling is particularly useful in the case of less
researched languages like Indian languages, for which
the most relevant features that affect the duration
pattern and the way they are inter-related have not
been studied in detail.
This paper is organized as follows. Section 2
gives the background for the work presented in this
paper. In Section 3, details about the speech corpus
that is used for the duration analysis are presented.
Section 4 describes the features considered for
duration modeling and subsequent generation of
feature vectors from which the CART based duration
model is trained. Section 5 describes the stepwise
construction of CART model for analysis on the
contribution and relative importance of various
features. In Section 6, objective evaluation of the
duration model, by root mean squared prediction error
and correlation between actual and predicted
durations, is presented.
RESEARCH ARTICLE OPEN ACCESS

2. Sangramsing Kayte Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 5, Issue 11, (Part - 4) November 2015, pp.115-119
www.ijera.com 116 | P a g e
II. Marathi TTS System
Marathi is an Indo-Aryan language spoken by about
71 million people mainly in the Indian state of
Maharashtra and neighboring states. Marathi is also
spoken in Israel and Mauritius. Marathi is thought to
be a descendent of Maharashtra, one of the Prakrit
languages which developed from Sanskrit. Marathi
first appeared in writing during the 11th century in
the form of inscriptions on stones and copper plates.
From the 13th century until the mid-20th century, it
was written with the Modi alphabet. Since 1950 it has
been written with the Devanāgarī alphabet[9,10].
The development of a high quality Marathi TTS
system [11] is under progress at HP Labs India. This
effort is a part of the Local Language Speech
Technology Initiative [7,11], which brings together
motivated groups around the world, providing tools,
expertise, support and training to enable TTS to be
developed in local languages. The aim of LLSTI is to
develop a TTS framework around Festival that will
allow for a rapid development of TTS in any
language[8,10,12].
2.1 Devanāgarī alphabet for Marathi
Vowels and vowel diacritics [9]
Consonants [9]

3. Sangramsing Kayte Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 5, Issue 11, (Part - 4) November 2015, pp.115-119
www.ijera.com 117 | P a g e
III. Speech corpus used
The present study of segmental durations in
natural speech is based on a corpus of around 90
minute duration, which consists of 1000 sentences
taken from three short stories. All the sentences are
spoken by a native Marathi male speaker in
expressive story reading style. The speaker is also a
professional radio artist. The recorded data is
manually segmented at phoneme level using
Pratt[13,14], thus yielding a total of 10,000 segments.
The data is divided randomly into training data
(10,000 segments, 85% of the total segments) and test
data (1000 segments, 15% of the total segments). A
total of 68 phonemes are analysed for their context-
dependent durations.
IV. Feature vector generation
Based on the literature [9,10,11,12,13], a number
of features are considered for segmental duration
prediction. Only those features that can be
automatically derived from text are considered. For
example, information about the focus or stress, accent
assignment and word boundary strength are not
considered even though they are known to affect
duration pattern. However, ‟stress‟ in Maharashtra
languages is not as clearly studied (both acoustically
and perceptually) as in a stress language like English.
Each segment in the corpus is annotated with the
following features together with the actual segment
(phoneme) duration:
Segment identity; e.g., /ka/, /ka/, /S/.
Segment features; e.g., vowel length, vowel height,
consonant type, consonant voicing.
Previous segment (immediate left context) features;
e.g., vowel length, vowel height, consonant type,
consonant voicing.
Next segment (immediate right context) features; e.g.,
vowel length, vowel height, consonant type,
consonant voicing.
Parent syllable structure; e.g., onset, coda, onset size,
coda size.
Position in the parent syllable; Position of the
segment in the syllable it is related to. The index
counts from 0.
Parent syllable initial; Returns 1 if the segment is the
first segment in the syllable it is part of, otherwise 0.
Parent syllable final; Returns 1 if the segment is the
last segment in the syllable it is part of, otherwise 0.
Parent syllable position type; the type of syllable
position in the word it is part of. This may be any of:
„single‟ for single syllable words, „initial‟ for word
initial syllables in a poly-syllabic word, „final‟ for
word final syllables in poly-syllabic words, and „mid‟
for syllables within poly-syllabic words. Number of
syllables in the parent word.
Position of the parent syllable; the position of the
syllable in the word it is part of. The index counts
from 0.
Parent syllables break information; Break level after
the parent syllable. This feature is categorical and it
has 4 possible values: 0 for word internal syllables, 1
for syllables occurring in word boundary, 3 for
syllables occurring in phrase boundary, 4 for syllables
occurring in sentence boundary.
 Phrase length in number of words.
 Position of phrase in the utterance.
 Number of phrases in the utterance.
The speech corpus used for modeling and analysis
is currently not optimal for duration modeling, since
we could not take care of to take care of data sparsity
problem or cover feature space. However, to reduce
the problem caused due to a small data set, care has
been taken to represent the feature space in a
generalized manner. For example, the segmental
context immediate left and right context is
represented using various features front vowel,
consonant type. instead of the absolute identities.
V. Generation of CART duration model
Classification and Regression Tree based
duration model is trained with feature data described
in Section 4. Since there is no previous knowledge
about the usefulness of the features and their relative
importance, CART‟s are built in a step-wise fashion
to establish the usefulness and relative importance of
the features. In this approach, each single feature is
taken in turn and a tree consisting of nodes
containing only the conditions imposed by that
feature is built. The single best tree is then kept and
each remaining feature is taken in turn and added to
the tree to find the best tree possible with just two
features. The procedure is then repeated for the third,
fourth, fifth feature and so on. This process continues
until no significant gain in accuracy is obtained by

4. Sangramsing Kayte Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 5, Issue 11, (Part - 4) November 2015, pp.115-119
www.ijera.com 118 | P a g e
adding more features. For running the CART
building process, ‟Wagon‟ classification and
regression tree tool [15,16] is used. Detailed analysis
on the usefulness of the proposed features and their
relative importance is given in Section 6.
5.1. Prediction of segmental duration
The segmental durations are predicted by traversing
the decision tree starting from the root node, taking
various paths satisfying the conditions at intermediate
nodes, till the leaf node is reached. The path taken
depends on various features like, the segment
identity, preceding and following segment identities,
position of the segment in parent syllable and
position of the syllable in parent word. The leaf node
contains the predicted value of segmental duration.
An example partial decision tree for segmental
duration prediction is shown in Figure 1. The tree
assigns different durations for segment /u/ when it
occurs in different contexts. A duration value of 110
ms is assigned when it satisfies the following criteria:
the preceding segment is /th/, parent syllable is the
final syllable in the parent word, and there is a break
(or pause) after the parent syllable. A duration value
of 70 ms is assigned when it satisfies the following
criteria: the preceding segment is /th/, parent syllable
is the final syllable in the parent word, and the parent
syllable is not at the end of a phrase break. A duration
value of 85 ms is assigned when the preceding
segment is /th/ and the following segment is /n/. A
duration value of 65 ms is assigned when the
preceding segment is /p/ and the following segment is
/d/.
VI. Objective evaluation and discussion
Objective evaluation of the duration models, by
root mean squared prediction error (RMSE) and
correlation between actual and predicted durations, is
performed. The duration model is trained with
training data (11282 segments, 90% of the total
segments) and evaluated with test data (1253
segments, 10% of the total segments).

5. Sangramsing Kayte Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 5, Issue 11, (Part - 4) November 2015, pp.115-119
www.ijera.com 119 | P a g e
Figure 1: An example partial decision tree (CART) for segmental duration prediction. The triangles depict
omitted parts.
Correlation obtained between the actual and predicted durations is 0.8997 and RMSE of prediction is 28.82 ms.
To assess the effectiveness of the features considered, CART‟s are built in a step-wise fashion (as described
in Section 5) and the results are shown in Table 1.
Table 1: Analysis on effectiveness of features in calculating segmental time.
Feature used Correlation
Segment Identity 0.6234
Next Segment (onset coda) 0.7112
Next Segment Vowel rounding (1 0) 0.7302
Next Segment Consonant Type 0.7423
Previous Syllable Break 0.7511
Syllable Coda Size 0.7587
Syllable Position (in word) 0.7691

6. Sangramsing Kayte Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 5, Issue 11, (Part - 4) November 2015, pp.115-119
www.ijera.com 120 | P a g e
Previous Segment Consonant Type 0.7744
Previous Segment Vowel Height 0.7869
Syllable Break 0.7987
The first column gives the names of the feature,
and the secondcolumn gives the
correlationobtainedbetween actual and predicted
durations by the addition of the successive features in
the CART modeling process. From the results, we
observe that the most important feature that
contributed to segmental duration prediction is the
identity of the segment itself. Other important
features in decreasing order of importance are:
Next segment‟s syllable structure - whether the
next segment is onset or coda of its parent syllable.
Next segment type - vowel rounding and consonant
type of next segment.
Previous syllable break information.
Parent syllable coda size.
Syllable position in word.
Previous segment type - consonant type and vowel
height of previous segment.
Parent syllable break information.
Though the segmental identity is the most
important feature, the prediction of segmental
duration is improved greatly by additional features
viz. syllable structure, immediate context type (right
and left) and syllable break information.
VII. Conclusions
Earliest results on data-driven Marathi duration
modeling is presented. Classification and Regression
Tree based approach for modeling segmental duration
is followed. A number of features are considered and
their usefulness and relative contribution to
segmental duration prediction is assessed. Work is in
progress on preparing large annotated speech corpora
and better prosody learning.
References
[1.] Monica Mundada, Sangramsing Kayte, Dr.
Bharti Gawali "Classification of Fluent and
Dysfluent Speech Using KNN Classifier"
International Journal of Advanced Research in
Computer Science and Software Engineering
Volume 4, Issue 9, September 2014
[2.] Monica Mundada, Bharti Gawali,
Sangramsing Kayte "Recognition and
classification of speech and its related fluency
disorders" International Journal of Computer
Science and Information Technologies
(IJCSIT)
[3.] Monica Mundada, Sangramsing Kayte
“Classification of speech and its related
fluency disorders Using KNN” ISSN2231-
0096 Volume-4 Number-3 Sept 2014
[4.] Sangramsing Kayte, Monica Mundada, Dr.
Charansing Kayte "Di-phone-Based
Concatenative Speech Synthesis System for
Hindi" International Journal of Advanced
Research in Computer Science and Software
Engineering -Volume 5, Issue 10, October-
2015
[5.] Sangramsing Kayte, Monica Mundada, Dr.
Charansing Kayte “Di-phone-Based
Concatenative Speech Synthesis Systems for
Marathi Language” OSR Journal of VLSI and
Signal Processing (IOSR-JVSP) Volume 5,
Issue 5, Ver. I (Sep –Oct. 2015), PP 76-81e-
ISSN: 2319 –4200, p-ISSN No. : 2319 –4197
www.iosrjournals.org
[6.] Hyunsong Chung andMark A. Huckvale,
“Linguistic factors affecting timing in Korean
with application to speech synthesis”, in
Eurospeech, Denmark, 2001.
[7.] Campbell, W., “Syllable-based Segmental
Durations”, In: G. Bailly, C. Beno it, and T.
Sawallis (Eds.), Talking machines: Theories,
models and designs, pp. 43-60, 1992.
[8.] Sangramsing Kayte, Monica Mundada, Dr.
Charansing Kayte "Implementation of Marathi
Language Speech Databases for Large
Dictionary" IOSR Journal of VLSI and Signal
Processing (IOSR-JVSP) Volume 5, Issue 6,
Ver. I (Nov -Dec. 2015), PP 40-45e-ISSN:
2319 –4200, p-ISSN No. : 2319 –4197
[9.] Sangramsing Kayte, Monica Mundada "Study
of Marathi Phones for Synthesis of Marathi
Speech from Text" International Journal of
Emerging Research in Management
&Technology ISSN: 2278-9359 (Volume-4,
Issue-10) October 2015.
[10.] Sangramsing Kayte, Monica Mundada, Dr.
Charansing Kayte "A Marathi Hidden-Markov
Model Based Speech Synthesis System" IOSR
Journal of VLSI and Signal Processing (IOSR-
JVSP) Volume 5, Issue 6, Ver. I (Nov -Dec.
2015), PP 34-39e-ISSN: 2319 –4200, p-ISSN
No. : 2319 –4197
[11.] Sangramsing Kayte, Monica Mundada, Dr.
Charansing Kayte " Performance Calculation
of Speech Synthesis Methods for Hindi
language IOSR Journal of VLSI and Signal
Processing (IOSR-JVSP) Volume 5, Issue 6,

7. Sangramsing Kayte Int. Journal of Engineering Research and Applications www.ijera.com
ISSN: 2248-9622, Vol. 5, Issue 11, (Part - 4) November 2015, pp.115-119
www.ijera.com 121 | P a g e
Ver. I (Nov -Dec. 2015), PP 13-19e-ISSN:
2319 –4200, p-ISSN No. : 2319 –4197
[12.] Sangramsing N.kayte “Marathi Isolated-
Word Automatic Speech Recognition System
based on Vector Quantization (VQ) approach”
101th Indian Science Congress Jammu
University 03th Feb to 07 Feb 2014.
[13.] Sangramsing Kayte, Dr. Bharti Gawali
“Marathi Speech Synthesis: A review”
International Journal on Recent and Innovation
Trends in Computing and Communication
ISSN: 2321-8169 Volume: 3 Issue: 6 3708 –
3711
[14.] Sangramsing Kayte, Monica Mundada,
Santosh Gaikwad, Bharti Gawali
"PERFORMANCE EVALUATION OF
SPEECH SYNTHESIS TECHNIQUES FOR
ENGLISH LANGUAGE " International
Congress on Information and Communication
Technology 9-10 October, 2015
[15.] Sangramsing Kayte, Monica Mundada, Dr.
Charansing Kayte "A Review of Unit
Selection Speech Synthesis International
Journal of Advanced Research in Computer
Science and Software Engineering -Volume 5,
Issue 10, October-2015
[16.] Taylor, P., R. Caley, and A.W. Black,“The
Edinburgh Speech Tools Library”,1.2.1
edition, University of
Edinburgh,http://www.cstr.ed.ac.uk/projects/sp
eechtools.html,2002.